Bluewin
@bluewin.bsky.social
Jack
Biophysics kinda guy for my PhD
Been teaching myself LLM research
It's quite fun!
But baking cookies is the most fun
Biophysics kinda guy for my PhD
Been teaching myself LLM research
It's quite fun!
But baking cookies is the most fun
So 1000 text generation tasks is about 2-10 minutes of breathing, running a laptop for 2 minutes, or driving 40 m. Or did I read that wrong?
Frankly this study feels a bit dishonest, every inference is scaled by 1k for no clear reason beyond bigger numbers in the graph if you don't read the legend
Frankly this study feels a bit dishonest, every inference is scaled by 1k for no clear reason beyond bigger numbers in the graph if you don't read the legend
November 22, 2024 at 4:10 AM
So 1000 text generation tasks is about 2-10 minutes of breathing, running a laptop for 2 minutes, or driving 40 m. Or did I read that wrong?
Frankly this study feels a bit dishonest, every inference is scaled by 1k for no clear reason beyond bigger numbers in the graph if you don't read the legend
Frankly this study feels a bit dishonest, every inference is scaled by 1k for no clear reason beyond bigger numbers in the graph if you don't read the legend
I made some little fellows. If you ask I will attempt to pick one that matches your vibe
November 20, 2024 at 2:00 PM
I made some little fellows. If you ask I will attempt to pick one that matches your vibe
I made a silly little game demo
It's not nearly done, I need better metrics for embedded space so I didn't need to use so many LLM calls.
But it's a story that can be altered, dynamically explored. Nothing too crazy yet, but I'm glad it's on board.
github.com/bluewin4/can...
It's not nearly done, I need better metrics for embedded space so I didn't need to use so many LLM calls.
But it's a story that can be altered, dynamically explored. Nothing too crazy yet, but I'm glad it's on board.
github.com/bluewin4/can...
November 16, 2024 at 10:44 AM
I made a silly little game demo
It's not nearly done, I need better metrics for embedded space so I didn't need to use so many LLM calls.
But it's a story that can be altered, dynamically explored. Nothing too crazy yet, but I'm glad it's on board.
github.com/bluewin4/can...
It's not nearly done, I need better metrics for embedded space so I didn't need to use so many LLM calls.
But it's a story that can be altered, dynamically explored. Nothing too crazy yet, but I'm glad it's on board.
github.com/bluewin4/can...
I made a cookie my boyfriend said looks like fried chicken
November 9, 2024 at 1:05 AM
I made a cookie my boyfriend said looks like fried chicken
If you see this, quote your green art 💚✨️
November 5, 2024 at 4:08 PM
If you see this, quote your green art 💚✨️
Look at this cookie dough
October 19, 2024 at 4:08 PM
Look at this cookie dough
I'm gonna just write about things I find cool
Like, did you know that slugs and snails can crawl over the edge of a razor sharp knife cause of their goo?
Like, did you know that slugs and snails can crawl over the edge of a razor sharp knife cause of their goo?
September 21, 2024 at 4:16 PM
I'm gonna just write about things I find cool
Like, did you know that slugs and snails can crawl over the edge of a razor sharp knife cause of their goo?
Like, did you know that slugs and snails can crawl over the edge of a razor sharp knife cause of their goo?
See this is why I can never do a rodent study
December 24, 2023 at 9:13 AM
See this is why I can never do a rodent study
Me an my bf sick as heck but full of good cheer
December 24, 2023 at 9:09 AM
Me an my bf sick as heck but full of good cheer
This paper explores how language models like GPT-3.5 and GPT-4 decline over time. The authors evaluated different versions of these models and found that their performance and behavior can vary greatly. This highlights the need for continuous monitoring of language models.
@alt-text.bsky.social
@alt-text.bsky.social
October 31, 2023 at 11:39 PM
This paper explores how language models like GPT-3.5 and GPT-4 decline over time. The authors evaluated different versions of these models and found that their performance and behavior can vary greatly. This highlights the need for continuous monitoring of language models.
@alt-text.bsky.social
@alt-text.bsky.social
They are also known for their ability to purify air. NASA found them particularly good, so their abundance is good for humans.
alt-text.bsky.social
alt-text.bsky.social
October 27, 2023 at 7:02 PM
They are also known for their ability to purify air. NASA found them particularly good, so their abundance is good for humans.
alt-text.bsky.social
alt-text.bsky.social
Does anyone know a good method for automatically generating alt-text for images? I keep forgetting and my pictures can be text heavy so it can be tedious.
I don't know how tagging works here yet so keyword blast: accessibility disability image model ♿ vision impaired
I don't know how tagging works here yet so keyword blast: accessibility disability image model ♿ vision impaired
October 27, 2023 at 5:13 PM
Does anyone know a good method for automatically generating alt-text for images? I keep forgetting and my pictures can be text heavy so it can be tedious.
I don't know how tagging works here yet so keyword blast: accessibility disability image model ♿ vision impaired
I don't know how tagging works here yet so keyword blast: accessibility disability image model ♿ vision impaired
This is the genetic algorithm they used.
I'm a little leery of just using the LLM as both the mutation source as well as the crossover, but it's still really cool.
I'm a little leery of just using the LLM as both the mutation source as well as the crossover, but it's still really cool.
![Genetic algorithm for implemented in evoprompting.
Please follow the instructions step by step to generate a better prompt.
1. Cross over the following prompts to generate a new prompt
Prompt 1: now you are a categorizer, your mission is to ascertain the sentiment of the provided text, either favorable or unfavorable.
Prompt 2: assign a sentiment label to the given sentence from ['negative, 'positive'] and return only the label without any other text.
2. Mutate the prompt generated in step 1 and generate a final prompt bracketed with <prompt> and </prompt>.
Response:
1. Crossover prompt: your mission is to ascertain the sentiment of the provided text and assign a sentiment label from ['negative, 'positive'].
2. <Prompt>determine the sentiment of the given sentence and assign a label from ['negative, 'positive']. </Prompt>](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:y34anx2p6co32wdkcu2tlxez/bafkreibyfjxg5lyuiu5rjsrfew6dbmkxva6js4ygxyulgtsq5ddullwtoi@jpeg)
October 27, 2023 at 5:09 PM
This is the genetic algorithm they used.
I'm a little leery of just using the LLM as both the mutation source as well as the crossover, but it's still really cool.
I'm a little leery of just using the LLM as both the mutation source as well as the crossover, but it's still really cool.
Okay I'm a sucker for evoprompting, it's a really cool combination of soft prompt tuning and selection pressures. Allowing for the model to optimize on two different information resivors simeltaneously.
This is the second paper and explains methodology in depth
arxiv.org/pdf/2309.085...
This is the second paper and explains methodology in depth
arxiv.org/pdf/2309.085...
October 27, 2023 at 5:01 PM
Okay I'm a sucker for evoprompting, it's a really cool combination of soft prompt tuning and selection pressures. Allowing for the model to optimize on two different information resivors simeltaneously.
This is the second paper and explains methodology in depth
arxiv.org/pdf/2309.085...
This is the second paper and explains methodology in depth
arxiv.org/pdf/2309.085...
Thinking about making these cookies again
October 23, 2023 at 11:58 PM
Thinking about making these cookies again
Blue art blue art! Post the art that's blue!
October 22, 2023 at 4:44 PM
Blue art blue art! Post the art that's blue!
Trying to pin down what the heck I'm doing with a bunch of lists of numbers and letters
October 22, 2023 at 12:11 PM
Trying to pin down what the heck I'm doing with a bunch of lists of numbers and letters
Been thinking about my favorite figure from a paper: Thermodynamics of Evolution and the Origin of Life
I really like how it endeavors to connect three disciplines using ideas of information and gives a grounding for people to complete interdisciplinary work.
I really like how it endeavors to connect three disciplines using ideas of information and gives a grounding for people to complete interdisciplinary work.
October 21, 2023 at 11:11 AM
Been thinking about my favorite figure from a paper: Thermodynamics of Evolution and the Origin of Life
I really like how it endeavors to connect three disciplines using ideas of information and gives a grounding for people to complete interdisciplinary work.
I really like how it endeavors to connect three disciplines using ideas of information and gives a grounding for people to complete interdisciplinary work.
Like the term or the concept? Because the term was in the 80's but the concept was a bit earlier when people realized there were "standard" protein folds.
October 20, 2023 at 10:52 PM
Like the term or the concept? Because the term was in the 80's but the concept was a bit earlier when people realized there were "standard" protein folds.